Handling of wind power forecast errors in the Nordic ...

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden – June 11-15, 2006

1

Handling of wind power forecast errors in the Nordic power market Hannele Holttinen Abstract-- Wind power production is variable by nature. These variations can be forecasted to some extent, by methods based on time series analysis or neural networks, for example. When forecasting more than 4 hours ahead, meteorological forecasts for wind speed are needed. The longer the time horizon, the more difficult it gets to forecast the production accurately, especially at hour-to-hour precision. Forecasting day ahead means having to correct on average 30-40 % of wind power production later. To handle the forecast errors, there can be trade up to the delivery hour, and the final imbalances to schedules are charged by regulating power prices. This paper outlines the forecast errors of wind power producers in the electricity market. The resulting benefits of shorter times between bids and delivery of production, as well as the benefits of pooling wind power production from larger areas are presented. Index Terms-- wind power; electricity market; forecast errors; forecasting; regulation market; imbalance pricing.

I. INTRODUCTION

W

ind energy is renewable, distributed generation characterised by large variations in the production. These variations smooth out to some extent when looking at a larger area [1,2]. The variations of wind power production occur in northern European latitudes due to meteorological weather systems passing the area, causing high winds, which calm down again. Forecasting wind power production relies on forecasted wind speeds in the area. Wind power forecasting day ahead is still new and the models are constantly subject to improvements [3]. Load forecasts have been studied for decades. However, it will not be possible to get to the same level of accuracy with wind power predictions as the load predictions are. Electricity consumption behaves with predictable diurnal and seasonal patterns, when looking at larger areas, with errors in the order of about 1.5–3 % of peak load, corresponding to an error of about 3–5 % of total energy, when forecasting day ahead. Wind power forecasts have a value both to power producers and to power systems. Day-ahead forecasts help the scheduling of conventional units: planning the start-ups and shut-downs of slow starting units in an optimised way and keeping the units running at best possible efficiency, saves fuel and thus operational costs of the power plants. Forecasts 1–2 hours ahead help keeping up the optimal amount of regulating capacity at the system operators’ use. However, both the system in question (production mix and load variations) and the properties of wind power production (correlation with load) have a strong effect on the results of how much benefit the improved predictions bring about [4]. Simulations of the England-Wales system show that the

prediction errors begin to affect the system fossil fuel costs when the wind power penetration is about 8 % (of yearly energy, 13 % of the capacity installed). At 20–30 % wind power penetrations (of energy), wind production forecasts can increase the savings in total fossil fuel costs by 13–35 % [5]. Today, the unit commitment and scheduling is done to a large extent by the electricity market: supply and demand bid to the market, which is settled at the most cost-effective way for each hour, day ahead. Also regulating power can be sold and bought at a market, closing an hour before, or even during the operating hour. All the producers with wind power in their generation mix, bidding to the market, need a forecast to base their bids on. Forecast errors result in supplying a different amount of energy than the bid, and if this is not corrected by the producer it will be penalized resulting in extra costs and thus reduced net income. The market design in this respect, that is how much the deviations of original bids to the market are penalised, can have a considerable effect on the wind power producer. Market design can also change the bidding strategy from simply minimising the error in energy [6,7]. Nordpool is the largest electricity market in Europe with longest history, since the beginning of the 90’s. It is working in the Nordic countries: Norway, Sweden, Finland and Denmark. In the spot market, hourly production can be traded. The market is cleared at noon, for the bids for the 24 hours the following day, 12–36 hours ahead. For Sweden, Finland and East Denmark, there exists also an intra-day market Elbas, which closes one hour before delivery, with continuous trade. Wind energy is traded in the market already today, by the Danish system operator. A part of wind energy production is bid in the spot market, thus avoiding the rescheduling of other production units in the area. Also some balance responsible players have some wind power production in their schedules in all four countries. This paper presents examples based on one year data for wind power predictions and actual, measured production. To quantify the benefits of operating in a shorter forecast horizon, the prediction errors are shown for different prediction horizons. To quantify the benefits of operating in a larger area for wind power production, the calculations are made by using simultaneous wind power data from the western and eastern parts of Denmark and one wind farm in Norway. II. FORECASTING WIND POWER PRODUCTION There are currently several different prediction approaches used in the models, most relying in different statistical

© Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden – June 11-15, 2006

2

III. REDUCTION OF PREDICTION ERRORS WITH SHORTER TIME

•

The proportion of produced energy that will be known x hours beforehand can be seen from Fig 1. Assuming the same level of wind power production ahead as presently (persistence), 90 % of wind power production will be known 1 hour beforehand. From the prediction model, 70 % of the wind power production will be known 9 hours before, 60 % 24 hours before and only 50 % 38 hours before. • For the Nordpool electricity market (prediction horizon 13...37 hours ahead), the mean absolute error (MAE) of wind power prediction is 8–9 % of installed capacity. However, for market operation it is relevant to know the error in the amount of energy produced, and this is 38 % of the yearly wind power production. For short time horizons, up till 3 hours, the persistence gives good results, even better than the WPPT. This is partly because there is no access on-line to the whole production of the area and in the WPPT model of year 2001 the up-scaling the production was not up-to-date. 100 % Absolute error, % of total production

methods like time series analysis and neural networks and some using physical description of the wind farm and topology as well [3]. The tool analysed in this paper is Wind Power Prediction Tool WPPT, developed in Denmark, since 1992. In 2001, the example year used in this paper, the version WPPT2 was used [8]. The model is based on statistical time series modelling taking as input the weather forecast for wind as well as the on-line measurements of wind power production for selected reference wind farms. The on-line measurements have negligible weight on prediction horizons of more than 12–18 hours. The model produces power production estimates for the reference wind farms, each representing a sub-area, and up-scales the production estimates for the sub-areas. Finally, the total prediction for the area is the sum of the predictions for sub-areas [8]. The predictions are made for 39 hours ahead and updated half hourly. However, the wind speed forecasts from the national weather service are obtained only 4 times a day. The resolution for the HIRLAM model was 17 km, and the forecast wind speed will be interpolated between the grid points for each of the 14 reference wind farms. The WPPT model is correcting the meteorological wind speed estimates for their tendency of producing larger wind speed values for longer time horizons as well as their lack of taking into account site specific diurnal variation.

90 % 80 % 70 % 60 % 50 % 40 % 30 %

Prediction model WPPT

20 %

Persistence

10 % 0% 0

HORIZON

Prediction errors for different prediction horizons were studied based on one year of operational data from West Denmark: wind power predictions as made by WPPT model during operation in 2001, and actual measured wind power production of West Denmark. One August week of missing predictions was excluded from the data. To see how much the prediction error increases with increasing forecast horizon, the predicted wind power production at different prediction horizons were compared to the actual production [9]. The results for the total production prediction errors of 1900 MW wind power are: • The correlation of predicted wind power production and actual wind power production keeps at a quite high level during the whole of the prediction horizon, above 0.90 for the first 12 hours and above 0.80 for up to 30 hours ahead (Nielsen & Madsen, 2000). Correlation tells us of the ability of the predictions to follow the ups and downs of the wind production. • When forecasting 6 hours ahead, the error was within ±100 MW for 61 % of the time. Large errors (> 500 MW) occurred nearly 1 % of the time. When forecasting 36 hours ahead, the errors were within ±100 MW 37 % of the time and large errors (outside ±500 MW) occurred 7 % of the time.

6

12

18

24

30

36

prediction horizon (hours)

Fig. 1. The sum of absolute prediction error for wind power predictions in 2001 for different prediction horizons, as a percentage of the total realised wind power production. Predictions from the model Wind Power Prediction Tool.

The forecast horizon is here taken from the constantly updated values of WPPT, however, the longer predictions are based on weather forecasts, which are only updated 4 times daily. It takes about 3 hours for the new forecast to come out in the WPPT output. The actual forecast horizon is thus 3 to 9 hours longer than stated here. Fig. 1 reveals the difficulty of acting at the market: even though the overall shape of the production curve can be predicted, the exact hourly value of wind power production is difficult to forecast 7… 38 hours ahead. This results in 30… 50 % of the total energy being forecasted wrongly. It has to be noted, however, that this is not the latest stateof-the-art of the forecasting models; improvements are expected in the future. The largest error component in the wind power production forecasts is the input from the weather forecast models. Meteorological institute weather service forecasts for wind speed and direction are not very accurate – partly because so far exact values at space and time have not been crucial for other applications. An accuracy of ±2 m/s and ±3 hours has been enough. For wind energy, however, this results in large errors in a day-ahead hourly market.

© Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden – June 11-15, 2006

3

Running the weather forecast models with several input values (ensemble forecasting) is used in latest developments, that gives information on the uncertainty of the wind speed forecast, and also help choose the right wind speed forecast as a basis for wind power predictions [10]. Getting better accuracy for weather forecasts for wind will improve the medium and long term (12… 36 hour) forecasts. Getting better knowledge of on-line wind power production in the area will improve both the short-term forecasts and the upscaling and estimation procedures of the statistical prediction model.

Denmark. Result for the total prediction error for the whole area was 1.47 TWh instead of 1.61 TWh just adding up the two. This means a 9 % reduction in the prediction error. If wind power production in East Denmark were the same as in West Denmark a 14 % reduction would happen in the prediction error when combining the two areas instead of calculating them separately. This just shows in theory the possibilities for error reduction, as in this case East Denmark is less in size and wind resource so equal wind power capacity is not realistic.

IV. REDUCTION OF PREDICTION ERROR IN A LARGER AREA Wind power prediction errors cancel out to some extent when the area is larger [11]. Making a production forecast to only one wind park results in more errors than making the forecast to tens or hundreds of wind parks covering a larger area. The same applies for load forecasting: predicting one load produces large errors compared to predicting the load in a larger area with hundreds of individual loads. A. Combining predictions in East and West Denmark The data was analysed to see the errors separately compared with the possibility of operating wind power in cooperation between West and East Denmark1. This is an example showing the smoothing effect of prediction errors in a larger area. In West Denmark, the wind parks have as a largest distance North-South less than 300 km and East-West less than 200 km (Fig.1). In the eastern part of Denmark, the wind parks are spread over area of 200 km (N-S) by 100 km (E-W) (excluding the Bornholm island). Including East Denmark adds 100 km or 50 % to West Denmark’s area, in the direction in which most weather systems pass (West–East) (Fig. 2). The installed wind power capacity in 2001 was about 550 MW in the East compared with 1900 MW in the West. Simultaneous prediction and production data were available from the system operators in Denmark in 2001, Elkraft System and Eltra. For predictions for Nordpool market predictions 12–36 h ahead, comparable data was available for a total of 8440 hours in 2001. The tool used for wind power prediction in East Denmark was developed at Elkraft. The key elements are essentially the same as already described for WPPT. Prediction errors for West and East Denmark have a weak correlation (0.4). For about a third of the time production is over predicted in the West and under predicted in the East, or vice versa, resulting in errors canceling each other out to some extent. The initial total prediction errors in a 12–36 hour market were 1.28 TWh for West Denmark and 0.33 TWh for East 1 The power systems of West and East Denmark are not connected. West Denmark is part of UCTE synchronous system and East Denmark is part of Nordel synchronous system. Both regions are interconnected to Germany and Sweden, and West Denmark also to Norway. The two system operators Eltra and Elkraft System merged to Energinet.dk in 2005.

Fig. 2. The area and transmission lines of Denmark: the western part is the Jutland peninsula and island Fyn, the largest islands on the eastern part are Zealand and Lolland.(Source:[12]).

B. Combining predictions in Norway and Denmark Wind power prediction for one site in Middle Norway was made with a Norwegian model with a different approach described more in detail in [13]. Simultaneous, historical data for predictions day-ahead and realised production in Denmark and Norway was available for 81 days in 2001. The data analysis showed that prediction errors in Norway and Denmark do not correlate (0.1). Prediction errors were to opposite directions about half of the time in Norway and Denmark. Prediction error decreases by roughly 25 % combining the predictions in Norway and Denmark, assuming similar amounts of wind power in each area. V. WIND POWER IN ELECTRICITY MARKETS Producers with wind power bidding on the electricity market need to forecast their wind power production. Through forecast, the wind power available can be estimated, selling all possible production. Forecast errors will result in an imbalance between delivered and bid energy. In balance settlement procedure, made by the system operators after the operating hour, the imbalances will be charged. In general the

© Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden – June 11-15, 2006

4

A. Reducing imbalance costs with shorter bidding horizon The market calculation is here made assuming different times between the bids and the delivery and a two-price model for charging the imbalance costs from the producers. Hourly wind power production and prediction data for year 2001 was used together with market price data from Nordpool: Elspot area price for West Denmark and up- and down-regulation prices for regulation market in West Denmark. The market calculation was made for three different bidding horizons: 13-37 hour forecasts for the current dayahead practice; 7-13 hour forecasts for a market clearing 4 times a day and finally continuous market with 2 hour forecasts. One hour was left for the operator to make the bids, this is why the forecasts were made one hour longer than the actual time between bid and delivery. For Nordpool Elspot the bids for the market have to be given before 12:00 the previous day. Forecast was calculated from the 11 o’clock prediction the previous day for hours 0:00–24:00 next day, and updated once a day. The predicted time series for a (theoretical) more flexible market (6–12 hours ahead) were calculated as 7… 13 hours ahead predictions updated four times a day. For the constantly operating market the information 2 hours before was used for the wind power prediction. The best prediction type here for year 2001 data was the persistence, using the realised wind power production 2 hours before as the estimate for delivery hour. The income from the market and the cost of regulation were calculated in the following way. Income I for the hour i is the predicted power ^Pi times Nordpool area price for West Denmark p spot

I i = Pˆ i ⋅ p spot

(1)

Cost c for the hour i is prediction error times regulation price preg. When wind power producer produces less than what has been bid to the market, the missing part will have to be purchased at up-regulation price, which is higher than the spot price received from the market. When wind power production is higher than the bid to the market, the surplus production is sold at down-regulation price, which is lower than the spot price, resulting in a negative cost in (2). Downregulation price can be negative, resulting in a cost instead of just lower income than if the prediction had been correct.

ci = ( Pˆ i − Pi ) ⋅ p reg

(2)

In Denmark, two-price model in the settlement of imbalances is used. This means that regulation price exists only for either up or down at each hour, depending on the direction of the system imbalance. Only when imbalances according to wind power prediction errors increase system imbalances, the regulation prices apply. When wind power

prediction errors are in the opposite direction –i.e. "help the system to balance" – imbalances are priced at Nordpool spot price. The imbalance of wind power was to the same direction as system imbalance about 70 % of time in 2001. For the remaining 30 % of time, wind power was being paid according to realised production. Finally, the net income is the income subtracted by costs, for the whole time period:

I TOTAL = ∑ I i − ci

(3)

i

The results are presented in Table 1. The Nordpool Elspot price for West Denmark during 2001 was on average 23.7 Euros/MWh. The average price for upregulation was 30.2 Euros/MWh and overproduction was rewarded on average only 12.3 Euros/MWh (downregulation price) (Fig. 3). Area price DK West, average 23.7 Eur/MWh

150

System price Nordpool, aver. 23.2 Eur/MWh

135

Regulation up, average 30.2 Eur/MWh

120

Regulation down, average 12.3 Eur/MWh

105 Eur/MWh

imbalances will be penalised and lead to reduced net income for the producer.

90 75 60 45 30 15 0 1

742

1483

2224

2965

3706

4447

5188

5929

6670

7411

8152

hour

Fig. 3. Market price data from Denmark West, year 2001, as duration curves. Regulation price exists only for either up or down for each hour. TABLE I INCOME AND COSTS FOR WIND POWER PRODUCER IN WESTERN DENMARK , WITH AND WITHOUT FORECASTS, CALCULATED FROM 2001 DATA 2.1.-16.8, 25.8.13-37 hour 7-13 hour 2 hour 31.12.2001 forecasts forecasts persistence Prediction error up/down as % of total 3.35 TWh 20 % / 19 % 15 % / 15 % 9%/9% Income Nordpool elspot, average Eur/MWh 22.9 22.9 22.8 Income Nordpool elspot, 22.4 22.4 22.5 predicted and realised production* average Eur/MWh Regulation: up/down 28 % / 25 % 37 % / 27 % 40 % / 29 % % of time 5%/5% 10 % / 11 % 15 % / 16 % % of energy 29.4 / 13.4 30.6 / 13.3 30.1 / 13.8 average Eur/MWh Regulation costs Eur/MWh regulated 5.9 5.2 3.8 Eur/MWh produced 2.3 1.5 0.7 Net income Nordpool average Eur/MWh 20.1 20.9 21.8 * this takes into account the 30 % of time when no regulation market price exists for wind power, as the imbalance is to opposite direction of system imbalance. During those hours the income is calculated from the realised production, not the predicted one.

If there were no forecast errors, the average price from Nordpool (area West Denmark) for wind power production would have been 22.9 Eur/MWh in 2001. This is somewhat less than the average price 23.7 Eur/MWh. For Nordpool system prices the realised wind power production would have received the average price. Wind power penetration is quite

© Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden – June 11-15, 2006

5

high in West Denmark, more than 16 % of gross demand in 2001, so it can be seen that wind power is affecting the price level in West Denmark resulting in less than average price [14]. For Nordpool 12–36 h market, prediction error for the year totals 0.68 TWh predicted too high and 0.67 TWh too low. This means that 39 % of the total yearly energy was predicted wrong. Taking into account that during some hours (about 30 % of time) the imbalance caused by wind power was to opposite direction than system imbalance, this results in 31 % of wind power production to be balanced at the regulation market. For a 6–12 h market, prediction error for the year totals 0.52 TWh predicted too high and 0.53 TWh too low. This means that 30 % of the total yearly energy would have been predicted wrong, and 21 % of the production had to be balanced at the regulation market. For a constantly operating hourly market, 18 % of the energy would be mispredicted, and 10 % of the production would have to be balanced at the regulation market. A more flexible market, allowing the bids for wind power to be updated 6–12 hours before, would reduce the regulation costs by 30 % and increase the net income by 4 % from 20.1 Eur/MWh to 20.9 Eur/MWh. An hourly operation, using persistence estimation from 2 hours before, would reduce the regulation costs for nearly 70 % and increase the net income by 8 %, to 21.8 Eur/MWh [9]. B. Imbalance pricing There are two issues relevant for the cost of imbalances. Both the amount of imbalance and how this imbalance is charged in the system influences the wind power producer. Reducing the forecast error of wind power means reducing the imbalances of wind power producers. For both the total wind power in power system and individual producers, this could be done by bidding closer to delivery or trading part of the imbalances in intra-day market closer to delivery. For individual producers, also aggregating the production over larger areas brings the total imbalance down, as individual imbalances tend to cancel each other out to some extent. For system operation, the knowledge of wind power forecasts can be derived either by making a prediction for wind power production in the whole system area, or by aggregating the information of all the wind power bids in the market, so basically aggregating wind power does not influence the net system imbalance that results in regulation needs. For the imbalance pricing, there is a difference in how the individual imbalances are charged from the producers. When the individual producer has been (partly) causing the total net system imbalance, the charges are the same in both one-price and two-price models: over or under production will be charged according to the regulation price for the hour in question. However, when the individual imbalance is to the opposite side of net system imbalance for the hour, this will either result in extra income (one-price-model) or will be reimbursed according to the spot price (two-price-model).

One-price model is in use in Norway and as the imbalance for wind power is about the same to both directions, this results in almost no extra regulation costs for wind power [15]. This means that there is no need to aggregate wind power production to reduce imbalance costs. In power systems where two-price model is in use, it is crucial for individual wind power producers to either form wind power pools or make a contract with larger conventional producers to reduce the imbalance costs. In two-price model it is also possible to reduce the imbalance costs by trading the imbalances up to one hour before the delivery hour. If a working intra-day market was at a wind power producer’s disposal, the correction of prediction errors could for a large part be traded at markets, instead of paying penalties for it. Elbas is a market like that, operating currently in Finland, Sweden and East Denmark. Taking the price series from Elbas market for year 2001, it was estimated how much the wind power producer would gain in this way [9]. With the predictions 2 hours ahead the producer trades the difference of the original bid and the more accurate prediction in Elbas market. For each hour there will be either a cost (from buying the missing production, at the highest realised Elbas price) or income (from selling the surplus production, at the lowest realised Elbas price). From 2001 data, there was slightly more buying than selling, so that the net cost was 1.6 Eur/MWh (per trading amount 1.2 TWh, cost is 0.6 Eur/MWh per the total wind power production). For regulating market, only 0.4 TWh needed to be adjusted, and the net income for wind power production was 21.5 Eur/MWh. This result shows that with a working intraday market used as an after sales tool, the net income for a wind power producer can be close to what it would be if the market was designed to be a short and flexible one (21.5 Eur/MWh compared with 1–2 hour market calculation 21.8 Eur/MWh in Table 1). However, looking from the power system point of view, it is not necessary to trade some amounts of wind power production back and forth, especially in a case where several individual wind power producers would try to reach the bid production amounts this way. VI. CONCLUSIONS AND DISCUSSION Wind power production, on an hourly level for 1–2 days ahead, is more difficult to predict than other production forms, or the load. The overall shape of the production curve can be predicted using weather forecasts and time series analysis. However, the peaks of wind power production are difficult to predict at hourly levels for both the exact amount and the exact occurrence in time. The benefits of predicting the production for larger areas was demonstrated combining the predictions for East and West Denmark. For 35 % of the time, the prediction errors for a 12–36 hour ahead market are to opposite directions. This resulted in a reduction of prediction error of 9 %. The prediction error would decrease more if the wind power

© Copyright KTH 2006

9th International Conference on Probabilistic Methods Applied to Power Systems KTH, Stockholm, Sweden – June 11-15, 2006

6

capacity would be more identical in the two areas –by simple up-scaling of the production in the East to the same level as in the West, a 14 % reduction in error would be achieved. Prediction errors were to opposite directions about half of the time in Middle Norway and Denmark and this would decrease prediction errors by roughly 25 % if similar amounts of wind power production was in place in Norway and Denmark. The benefits of shortening the forecast horizon was demonstrated for West Denmark data. The errors amounted to 50 % of being forecasted wrong, when forecasting the exact hourly value of wind power production 38 hours ahead. The figures where 30 % of the total energy 7 hours ahead and 10 % of the total energy 1 hour ahead, respectively. It has to be noted, however, that this was for 2001 data, not the latest state-of-the-art of the forecasting models. The benefits of shortening the forecast horizon was further demonstrated by electricity market calculations. A more flexible market, allowing the bids for wind power to be updated 4 times daily, with predictions of 6–12 hours ahead, would reduce the regulation costs for 30 % and increase the net income by 4 %. Hourly operation would reduce the regulation costs for 70 % and increase the net income by 8 %. Using an after sales tool like Elbas for trading the estimated surplus or missing production 2 hours before delivery would reduce the regulation costs by 70 % and increase the net income by 7 %. The results are based on year 2001 data of West Denmark, where wind power penetration is considerable and can be seen to influence the prices. The assumption has been made, that the same price level would apply when shortening the time between bids and delivery, not taking into account the implications of a shorter market to other production forms and actors. For Elbas after sales prices, no impact of wind power production or bottlenecks to the price level has been assumed. If the price level at regulating market was higher in penalising the imbalances, the benefit for a flexible market, or after-sales tool, could be greater. On the other hand, acting at flexible markets could also bring about extra trading costs. For a wind power producer, selling his production at a market, there is a clear benefit for trading as close to the delivery as possible, because this reduces the prediction errors and thus extra costs from imbalances. Market design has a strong influence on wind power. In two-price models, forming wind power pools that can forecast for a larger area reduces the imbalance costs. For day-ahead markets, also trading part of the imbalances in intra-day markets can reduce imbalance costs. For the power system, all imbalances do not need to be balanced one-by-one, only the net imbalance. In a large system this results in considerable benefit, when most of the individual imbalances counteract one another. This should be reflected by the balance settlement rules as well.

VII. ACKNOWLEDGMENTS The author gratefully acknowledges the use of prediction data from year 2001 from energinet.dk (former Eltra and Elkraft System) and John Bjornar Bremnes at Norwegian Meteorological Institute. VIII. REFERENCES [1]

[2] [3]

[4]

[5]

[6]

[7]

[8]

[9] [10]

[11]

[12]

[13]

[14]

[15]

Giebel, G, 2001. On the Benefits of Distributed Generation of Wind Energy in Europe. PhD thesis from the Carl von Ossietzky Universität Oldenburg. VDI-Verlag, Schriftenreihe Energietechnik, 2001. ISBN 3-18344406-2 Holttinen, H. 2005a. Hourly wind power variations in the Nordic countries. Wind Energy, vol. 8, 2, ss. 173 - 195 Giebel, G, Brownsword, R, Kariniotakis, G, 2003. The State-of-the-art in Short-term prediction of wind power. A literature overview. EU project ANEMOS (ENK5-CT-2002-00665). Available at http://anemos.cma.fr/download/ANEMOS_D1.1_StateOfTheArt_v1.1.pdf Milligan, M, Miller, A H, Chapman, F, 1995. Estimating the economic value of wind forecasting to utilities. Proceedings of Windpower’95, Washington D.C, March 27-30, 1995. A report NREL/TP-441-7803 available at http://www.nrel.gov/publications/ Watson, S J, Giebel, G, Joensen, A, 1999. The economic value of accurate wind power forecasting to utilities. Proceedings of European Wind Energy Conference, 1-5 March 1999, Nice, France, pp 1009-1012. Bathurst, G N, Weatherill, J, Strbac, G, 2002. Trading wind generation in short term energy markets. IEEE Transactions on Power Systems, Vol. 17, No 3, August, 2002. Nielsen, C S, Ravn, H F, 2003. Criteria in Short-term Wind Power Prognosis. CD-Rom Proceedings of the European Wind Energy Conference & Exhibition EWEC 2003, June 16-19, 2003, Madrid, Spain. Nielsen, T S, Madsen, H, 2000. WPPT –a tool for wind power prediction. In proceedings of EWEA special topic conference, 25 – 27th September, 2000 in Kassel Holttinen, H., 2005. Optimal electricity market for wind power. Energy Policy, vol. 33, 16, ss. 2052 - 2063. Giebel, G (ed.), Badger, J, Landberg, L, Nielsen, H Aa, Nielsen, TS, Madsen, H, Sattler, K, Feddersen, H, Vedel, H, Tøfting, J, Kruse, L, Voulund, L, 2005. Wind power prediction using ensembles. Risø-R1527(EN), 43 p. Focken, U, Lange, M and Waldl, H M, 2001. Previento – A Wind Power Prediction System with an Innovative Upscaling Algorighm. In proceedings of European Wind Energy Conference, 1.-5th July, 2001, Copenhagen. Hilger, C, 2002. Wind Power and Impacts on System Operations. CIGRE Int. Workshop on Wind Power and the Impacts on Power Systems, 17.-18 June, 2002, Oslo, Norway. Bremnes, J.B., Villanger, F., 2002. Probablistic forecasts for daily power production. In Proceedings of Global Wind Power Conference, Paris, 1.-4 th April, 2002. Holttinen, H, 2004. The impact of large scale wind power production on the Nordic electricity system. Espoo, VTT Processes. 82 p. + app. 111 p. VTT Publications; 554 Available at http://www.vtt.fi/inf/pdf/publications/2004/P554.pdf Gustafsson, M, 2002. Wind power, generation imbalances and regulation cost. Presented at CIGRE. June 2002, Oslo, Norway.

IX. BIOGRAPHY Hannele Holttinen (MSc’1991, PhD’2004) was born in in Helsinki, Finland in 1964. She studied at the Helsinki Tehcnical University. Her Master’s Thesis was on turbulence modeling and her PhD on power system impacts of wind energy. She has worked for VTT Technical Research Centre of Finland since 1989 in different fields of wind energy research including resource assessment and measurements, production and failure statistics, offshore and arctic wind power feasibility. Since 2000, her main special fields of interest involve wind power impacts on power systems and electricity markets. .

© Copyright KTH 2006